AI4Mat - NeurIPS 2023

AI for

Accelerated Materials Design

NeurIPS '22 Workshop

AI4Mat NeurIPS 2022 Workshop Recap

On December 2, 2022, the AI4Mat Program Committee hosted the 1st workshop on AI for Accelerated Design (AI4Mat). Our goal was to bring together researchers and domain experts from both AI and materials science together to discuss interdisciplinary technical challenges in building towards the vision of Self-Driving Materials Laboratories. Overall, we believe that we achieved many of the goals of the workshop pertaining to productive interdisciplinary discussion leading to new insights and also discovered promising future directions across the different technical themes.

Snapshot of AI4Mat Workshop on December 2nd, 2022

General Recap

Schedule Overview of the AI4Mat Workshop on December 2nd, 2022

The schematic schedule above shows the general outline of the workshop and its breakdown along the three major themes: 1. AI-Guided Design; 2. Automated Chemical Synthesis; 3. Automated Material Characterization. The opening remarks provided a general introduction to the workshop themes, schedule and speakers and highlighted some statistics related to the scientific contributions:

Workshop Contributions Along Scientific Categories

Workshop Contributions Along Workshop Tracks

As shown in the figures above, the contributions skewed heavily towards AI-Guided Design, which more closely aligns with paper submissions to the primary NeurIPS conference. One of the main goals of the workshop is to further promote, encourage and connect researchers who are working in currently underrepresented themes of Automated Chemical Synthesis and Automated Material Characterization. Increased research activity in these areas will drive the community closer towards the realization of closed-loop AI-infused automated materials design with meaningful impact on real-world applications. Automated Chemical Synthesis and Automated Material Characterization often require real-world tools and equipment, which can often create a greater barrier to entry for researchers to make meaningful progress. The contribution statistics also show that the paper category generally dominated the submissions, which also partially reflects the bias of NeurIPS. In the future, we hope that the distribution of submission will be less heavily skewed towards particular themes and categories given that many different pieces are needed to productively work towards the greater vision. While a general bias given the general focus of NeurIPS is expected and it’s worthwhile to continue to highlight advancement in AI-Guided Design, we believe that the power of advanced machine learning is currently underutilized in Automated Chemical Synthesis and Automated Material Characterization, prompting us to highlight and introduce cross-disciplinary works and speakers in those disciplines.

Geographic Distribution of Authors

Geographic Distribution of Reviewers

The program committee reviewed the geographical locations of contributing authors and reviewers. The author bias skews towards North America, Europe and East Asia with the reviewer bias skewing heavily towards North America. In the future, one potential improvement would be to include greater representation of the reviewer population across the geographies of contributing authors. Additionally, future workshops should encourage submissions from institutions in underrepresented areas in the southern hemisphere to increase the global diversity of the contributing and reviewing population.

Panel Discussion

The workshop began with a cross-disciplinary panel spanning the technical themes of the workshop, as well as representation from academic, governmental and industrial research organization.

Schedule Overview of the AI4Mat Workshop on December 2nd, 2022

The panel discussion initially focused on providing thoughts and advice for active and interested researchers building their careers in the highly interdisciplinary and complex field of AI and materials science, followed by an extensive discussion of some of the ongoing challenges in the field. Some of the major takeaways from the panel include:

The panel generally recommended new and active researchers to build expertise in 1-2 core areas of their preferences and leverage interdisciplinary networking and teamwork to perform meaningful research work at the intersection of AI and materials science.
AI is already being applied in targeted, consequential ways in real-world materials design workflows and applications. Major breakthroughs that fulfill the ambitious goals outlined in the workshop, however, are still many years away creating opportunities for meaningful research work.
The fragmented nature of data collection and dataset usage is a major challenge in scaling AI for automated materials design. As systems and technical ambitions become more and more complex, the ability to manage different kinds of data at larger scales becomes more and more important. This can also greatly affect reproducibility, especially in experimental settings with complex equipment.

AI4Mat Panel Session

Keynotes and Spotlight Presentations

The workshop presented keynotes and spotlight talks from the three thematic tracks along with interactive discussion and a poster session showcasing all contributions.

Schedule Overview of the AI4Mat Workshop on December 2nd, 2022

The keynotes showcased real-world use cases of advanced AI in materials design, including natural language processing for determining fabrication paths of new materials systems, building AI-based tools to enable robotic synthesis of molecular materials and the application of deep learning for advanced electron microscopy at state-of-the-art atomic scales. Some major takeaways from the various keynotes and spotlights include:

AI-Guided Design:
- AI with some human action in the loop can already significantly accelerate materials discovery in real-world use cases, including conductive thin films and 3D printed bone replacements.
- AI can significantly speed up the discovery process by effectively using compute in approximating simulations and doing effective searching in large design spaces.
Automated Chemical Synthesis:
- Designing AI systems with actionable chemical steps in mind can significantly improve the synthesizability of generated structures and thereby help optimize and automate various kinds of materials fabrication procedures.
- Sample efficiency in experimental materials fabrication is a key challenge that AI can help with early closed-loop systems already showing promising results.
Automated Material Characterization:
- Data scarcity remains a challenge given the cost of obtaining high-quality characterization data and associated labels in characterization settings. This creates a prominent opportunity for generating high-quality synthetic data using advanced techniques and data augmentation techniques.
- Interpretability is a major consideration in characterization to continue to enable researchers to make quicker progress in materials analysis that will help inform subsequent materials design decisions.

Poster Session

The poster session and technical discussions of the workshop also enabled new connections to be built for interested researchers across international academic, government and industrial institutions. The set of accepted works can be found on the website and we hope interested researchers continue to connect and spark meaningful research going forward.

AI4Mat Poster Session

Future Directions

The 1st AI for Accelerated Materials Design workshop at NeurIPS 2022 achieved its primary objective in showcasing research and bringing together domain experts from both AI and materials science. In doing this, several insights and future directions emerged through diverse sets of discussion that aim to drive research in the community forward. Here are some of the goals the program committee believes warrant greater focus by the research community:

Enhance the representation of research work in underrepresented technical areas (Automated Chemical Synthesis, Automated Material Characterization) and geographic areas of contributing authors and reviewers.
Continue discussion and research towards a more unified approach for recording and sharing data across different modalities, including simulation and experimental settings across different sets of equipment and relevant metadata.
Continue to encourage and highlight AI-infusion and cross-disciplinary for consequential challenges in automated materials design, such as enhancing sample efficiency for automated material design targets, providing high quality data for data-scarce characterization problems and advanced algorithm development for inspiring materials design cases.